Point to Space: Data-driven Soil Moisture Spatial Mapping using Machine-Learning for Small Catchments in Western Germany

Soil moisture is a crucial variable in the earths critical zone. It depends on multiple factors such as climate, topographic conditions, soil characteristics and affects energy and water fluxes across the land-atmosphere interface and, therefore, is highly important for terrestrial ecosystems, ecosystem management and agriculture. The accurate mapping of soil moisture across time and space is challenging but highly desirable.

One option is to deploy ground-based moisture sensors at the point-scale and to interpolate and/or map the measurements into space. We have developed a data-driven approach to map the soil moisture in a reference area from point measurements at specific time points and the covariates: location, topographic conditions and soil characteristics. We tested the mapping capacity of different two machine-learning algorithms (Random Forest and Neural Networks) and compared those with Ordinary Kriging as standard method. Our questions were: 1) How accurate are the machine-learning methods for soil moisture mapping? 2) Which covariates are most important? 3) How does mapping accuracy vary with data density and temporal resolution? 

We used soil moisture data from the TERENO experimental sites Wüstebach and Rollesbroich located in western Germany. These small catchments are equipped with a dense network of soil moisture sensors using time domain reflectometry (TDR) that has been operated since 2010 (Bogena et al., 2010; Zacharias et al., 2011). From this, we created 2700 point-based soil moisture data sets at specific time points, specific depth and for various numbers of sensor locations. Then we merged these data sets with sampled data on soil texture and chemical composition (Qu et al., 2016; Gottselig, et al., 2017; ) as well as remote sensed terrain data. These time stamp specific point-based soil moisture measurements were mapped using Ordinary Kriging (OK), Random Forest (RF) and Neural Networks (ANN) using combinations of the soil and terrain attributes as well as geometric distances between sensor locations as covariates. Each model was trained (80% subset) and tested (20% subset) on the point-based data sets. 

In general, average model accuracy across the methods and individual data set types (depth, number of sensor locations, temporal averaging) was relatively low with R2 values of approximately 0.2-0.5. This originated in the high variability of soil moisture. Surprisingly, models using the spatial structure of the domain (using distances between sensors as covariates) already yield an R2 of approximately 0.45. Further adding covariates such as soil and terrain attributes did not substantially improve the accuracy for these models. In comparison,using only terrain attributes as covariates for RF and ANN did yield an accuracy of R2 of 0.25-0.27.The trained models were then used to map soil moisture onto the entire study area. This resulted in maps with interesting patterns that differed between the individual methods—even when using same covariate types.

Finally, it can be concluded that for spatial interpolation of soil moisture the Random Forest algorithm using distance between sensor locations as covariates is a promising alternative to Ordinary Kriging from the point of accuracy and simplicity.